PREDICTING OPTIMAL LEE FILTER WINDOW SIZE FOR SENTINEL-1 SAR IMAGES USING TRANSFER LEARNING ON MOBILENETV2

Abstract

A method for a priori selection of the optimal Lee filter window size for speckle noise suppression in Sentinel-1 SAR images is presented. The task is formulated as a direct multiclass classification problem. The proposed model employs transfer learning with a compact MobileNetV2 backbone adapted to single-channel SAR input and equipped with a lightweight classification head. The training corpus is generated synthetically: Sentinel-2 patches are aligned to the Sentinel-1 distribution via histogram matching and subjected to dynamic multiplicative noise with gamma distribution and controlled spatial correlation. Ground truth labels are determined by exhaustive Lee filtering with window sizes of 3×3, 5×5, 7×7, 9×9, and 11×11, selecting argmax over the full quality metric. Six independent experiments were conducted for PSNR, SSIM, MS-SSIM, FSIM, HaarPSI, and MDSI criteria. On a test set of 382 patches (1024×1024), the best overall accuracy reached 87.17% (FSIM), balanced accuracy achieved 88.94% (MS-SSIM), minimum calibration error ECE was 0.0371 (PSNR), Brier score was 0.0359 (FSIM), with Top-2 accuracy in the range of 99.21–100.00% (100.00% for PSNR). Across ENL bins, accuracy remained  and reached 91.1% for  with PSNR and FSIM; across speckle spatial frequency scale, peaks of 93.7% (FSIM) at  and 93.0% (MS-SSIM) at  were observed. Critically, the method does not require a priori knowledge of the equivalent number of looks (ENL), which is consistent with accuracy stability in the range . The method is suitable for operational integration into production SAR pipelines for preprocessing and computational resource planning.